Qwen3.8-27B: 32 demos put together in one sitting

Alongside the release of the Qwen3.8-27B language model came a collection of demonstration pages built entirely by the model itself. The author of the collection stresses two things: all 32 demos were generated “in one sitting” on release day, and every card leads to its own separate, self-contained page.

The genres in the gallery vary noticeably. There are graphical experiments — a voxel garden with pagodas, terrain that “breathes” through color based on fractal Brownian motion, shader metal droplets, an organic blob shimmering with an oily film, and inky tentacles crawling across the page. The simulations stand apart: a cellular automaton of a forest fire (including a “chalkboard” version), a growing coral reef with schools of fish, and the spread of an infection through a hand-drawn little town.

It wasn't without arcade-style games — racing behind the wheel through a horde of zombies, a “battle royale” of sixty AI fighters squeezed by a shrinking storm circle. And there's a whole series of design exercises too: sites in brutalist, risograph and Swiss-typography styles, a monochrome homage to the 1995 Virtual Boy console, a retail stationery catalog, a “terminal” SaaS-service price list, and an ASCII zine about exit codes and dead ends.

The technical side of the collection is intriguing. Some of the demos are duplicated in several variants — a base one, NVFP4 and Q8. In essence, these are the same pages (a dithered bento layout, the stationery catalog, the Virtual Boy microsite, the hand-drawn “moss notebook,” the voxel garden), but generated at different compute precisions. Such a comparison makes it possible to clearly assess how quantization and number format affect the result the model produces.

In form, this is neither a benchmark nor a report but a showcase of capabilities: a set of small yet finished frontend projects demonstrating what the model is able to produce on its very first day of availability. The most useful thing here is precisely the parallel versions at different precisions — a rare case where you can compare the same task carried out by a single model under different output settings.

Source: Lobsters